What problem it solves

Rust developers need host-side access to the CUDA runtime and tools for expressing GPU kernels while preserving Rust's ownership and borrowing rules. The platform provides separate programming models for workloads that require direct thread-level control and workloads naturally expressed through tensor operations.

Key features

• Shared CUDA runtime — Host-side crates provide driver bindings, contexts, streams, device buffers, events, and lazy composable device operations.

  Source: README, cuda-core, cuda-async

• SIMT kernel programming — The cuda-oxide component includes a custom rustc code-generation backend, a Cargo helper, and crates for targeting the traditional CUDA SIMT kernel programming model with direct control over device-side execution.

  Source: README, cuda-oxide

• Tile kernel programming — The cutile-rs component provides a proc-macro plugin and crates for writing kernels with the CUDA Tile model, where tensors are partitioned into sub-tensors and kernels operate on whole tiles.

  Source: README, cutile-rs

• Ownership across launches — Mutable buffers are divided into disjoint pieces before a launch, immutable buffers can be shared, and launchers retain the relevant borrows while GPU work remains in flight.

  Source: README, cuda-repr

• Interoperable kernel models — Tile and SIMT kernels can interoperate on the same stream over shared device buffers, allowing an application to use both models.

  Source: README

Who it may suit

Inference: this may suit Rust developers who need CUDA runtime integration, direct control over GPU threads, or tensor-oriented kernels. It may also suit teams evaluating how Rust ownership rules can be carried across GPU launch boundaries.

Topics to discuss

• Which kernel programming model should a new CUDA Rust workload use: Tile, SIMT, or both?

• How should the shared runtime evolve as Tile and SIMT components continue to mature?

• What examples, reusable kernels, and documentation would be most useful for CUDA Rust users?

Written by ReadmeX AI from the project's README, source and public GitHub data; it may contain mistakes.

Read the original README https://github.com/NVIDIA/cuda-rust#readme

ReadmeX opened this community for a trending project; it is not run by the project. Maintainers can claim it from the community page.